Research On Channel Decoding Algorithm Based On Deep Learning

Since 2006,the emergence of effective deep network training methods has made deep learn-ing encounter the third climax of its development in this century.The success of deep learning in speech recognition and image classification in 2012 enabled deep learning research to rise again while gradually extending to more fields.Since 2016,the intelligent communication based on deep learn-ing techonology has begun to show explosive development which basically covers all aspects of the physical layer.To this article,it will mainly focus on the repspect of learning in channel decoding algorithms,and propose enhanced decoding schemes based on deep learning for LDPC codes,polar codes in the 5G standard and Turbo codes in 4G standards.Specifically,we will comprehensively study model-driven decoding schemes and data-driven decoding schemes and give specific and ef-fective training methods.First,we introduce some basic concepts of LDPC codes,and briefly conclude the quasi-cyclic(QC)extension characteristics of 5G-LDPC codes based on the prototype graph extension.Then we use the deep unfolding method to achieve the decoding network which is called Deep learning-based min-sum(D-LMS)decoding algorithm based on the QC extension characteristics.It also explains the methods and reasons of specific design details of the nework structures and hyperparameters.After that,we get the enhanced algorithm based on the K-sparse layer in order to further improve the decoding perfor-mance through K-sparse training and secondary OSD decoding algorithm(D-LMS-Ksparse-OSD).While introducing hyperparameter settings,this article gives a data set optimization algorithm that can further accelerate network convergence.Finally,simulation results verify that the performance of D-LMS algorithm is better than MS algorithm and the existing learning-based LAMS algorithm under different code rates and code lengths.At the same time,D-LMS-Ksparse-OSD achieves an enhanced performance than the D-LMS under short codes.Next,we explore how to implement a polar code decoding algorithm with a decoding structure similar to previous chapter.Firstly,the basic concepts related to polarization codes are briefly men-tioned which includes the process of channel combining and splitting.Then,it leads to the concept of channel polarization and the method of polar code construction.After that,we introduce two classic decoding algorithms of the polarization codes in detail,the serial cancellation(Successive Cancel-lation,SC)and the serial cancellation list(SC List)algorithm.To realize the purpose of sharing the similar decoding structure with LDPC,we combine the D-LMS algorithm with the sparse graph method to propose the Sparse-D-LMS algorithm while providing an enhancd decoding method which is called noise assisted D-LMS list(Na-Sparse-D-LMSL)algorithm.The simulation results show that the Sparse-D-LMS algorithm can achieve similar performance to Sparse-BP with lower complexity and do not have to estimate the noise variance.The performance of the Na-Sparse-D-LMSL algo-rithm is equivalent to that of the SCL algorithm.Although its performance is inferior to CA-SCL,it has a higher degree of parallelism and has a similar decoding structure to LDPC decoding.Finally,we introduce the NN-Turbo decoder based on a mixed driven method.At the begining,we introduce the encoding methods of Turbo code under LTE standard which includes specific trans-fer function,tail bit generation rules,and QPP interleaving methods.Then we combine the related knowledge of hidden Markov process to derive the MAP-based BCJR decoding method used in the component decoder in detail.After that,we achieve the specific structure of NN-Turbo on the basis of unfolding normal Turbo structure,the core of the component decoder is replaced with a data-driven network.Then we analyze the advantages and disadvantages of two types of network cores: Recurrent Neural Network(RNN)and Convolutional Neural Network(CNN).In order to solve the problem that the data-driven core is difficult to train,a network training method based on gradually increasing the entropy of the data set is given.Meanwhile,a NN-Turbo training scheme compatible with different code lengths and different code rates is given.The results show that RNN is most suitable as the core of NN-Turbo component decoder,and the performance of NN-Turbo is better than Max-Log-MAP and can achieve similar performance to Log-MAP decoder without noise estimation.The code length compatibility can be achieved through the recurrent character of the RNN network,and the code rate compatibility can be achieved through transfer learning.The simulation results show that there is only a small performance loss.